A phenotypic drug discovery approach by latent interaction in deep learning

Area of Science:

• Computational biology
• Drug discovery
• Bioinformatics

Background:

• Current drug discovery relies on binding assays, which overlook complex molecular interactions like cis-regulatory elements.
• This limitation hinders computational method development and drug discovery potential.
• Higher-order interactions and data scarcity are significant challenges in traditional approaches.

Purpose of the Study:

• To develop a deep learning model for drug discovery using an end-to-end approach based solely on therapeutic drug information.
• To overcome limitations of binding assays and insufficient binding specificity data.
• To address challenges of data scarcity and complex molecular interactions in drug discovery.

Main Methods:

• Developed a deep learning model transforming textual drug and virus genetic information into high-dimensional latent representations.
• Employed an end-to-end approach relying exclusively on therapeutic drug information.
• Utilized various modeling skills and data augmentation techniques to handle data scarcity and complex interactions.

Main Results:

• The model implicitly considers epistasis and chemical-genetic interactions, outperforming traditional methods.
• Demonstrated outstanding out-of-sample validation performance, even with unknown complex interactions.
• Highlighted the importance of chemical diversity in model training for drug discovery.

Conclusions:

• Deep learning offers a feasible approach for drug discovery in data-scarce scenarios.
• The developed model provides a promising alternative for drug discovery when underlying mechanisms are not fully understood.
• This method effectively bypasses the need for detailed binding specificities, addressing a key limitation in computational drug discovery.